Article
Improved Prediction of Stream Flow Based on Updating Land Cover Maps with Remotely Sensed Forest Change Detection Alexander J. Hernandez 1, *, Sean P. Healey 2 , Hongsheng Huang 3 and R. Douglas Ramsey 1 1 2 3
*
Wildland Resources Department, Utah State University, Logan, UT 84322-5230, USA;
[email protected] USDA Forest Service, Rocky Mountain Research Station, 507 25th Street, Ogden, UT 84401, USA;
[email protected] Key Laboratory of Poyang Lake Basin Agricultural Resources and Ecology of Jiangxi Province, Jiangxi Agricultural University, Nanchang 330045, China;
[email protected] Correspondence:
[email protected]; Tel.: +1-435-797-2572
Received: 26 April 2018; Accepted: 31 May 2018; Published: 2 June 2018
Abstract: The water balance in a watershed can be disrupted by forest disturbances such as harvests and fires. Techniques to accurately and efficiently map forest cover changes due to disturbance are evolving quickly, and it is of interest to ask how useful maps of different types of disturbances over time can be in the prediction of water yield. We assessed the benefits of using land cover maps produced at annual vs. five-year intervals in the prediction of monthly streamflows across 10 watersheds contained entirely within the US National Forest System. We found that annually updating land cover maps with forest disturbance data significantly improved water yield predictions using the Soil and Water Assessment Tool (SWAT; p < 0.01 improvement for both the Nash–Sutcliffe efficiency measure and the ratio of the root mean square error to the standard deviation of the measured data). Improvement related to using annually updated land cover maps was directly related to the amount of disturbance observed in a watershed. Our results lay a foundation to apply new high-resolution disturbance datasets in the field of hydrologic modeling to monitor ungauged watersheds and to explore potential water yield changes in watersheds if climate conditions or management practices were to change forest disturbance processes. Keywords: forest disturbance; time series; SWAT; water dynamics; land cover
1. Introduction Changes in forest structure can impact water balance dynamics of a given catchment area. There is ample evidence that disturbance—such as fires [1], harvests [2], and bark beetles outbreaks [3]—can affect interception, infiltration, evapotranspiration, and runoff. However, while a variety of options exist for using remote sensing platforms such as Landsat to develop long-term spatial records of disturbance [4], there has been no effort to date that the authors know of to integrate these products in the models of annual stream flow. Approximately 65% of the water supply in the western US states originates on forested lands, with 33% of the total being contributed solely from upper watersheds located within national forests and grasslands [5]. Prediction of forest water yield, particularly as it is affected by recent disturbances, is therefore important in managing the water supply of millions of people. While land cover change information has been included in hydrologic models to better understand water dynamics [6–9], no reference was found of work that integrates comprehensive, annually specific disturbance maps into hydrologic models. There are, however, examples [10–12] of recent research
Forests 2018, 9, 317; doi:10.3390/f9060317
www.mdpi.com/journal/forests
Forests 2018, 9, 317
2 of 19
using different dates (five years apart) of the National Land Cover Dataset NLCD [13,14] in an effort to integrate land cover change in hydrologic simulations. The current research was conducted to address this question: do annual land cover maps, updated annually to reflect forest disturbances, produce better water yield estimates compared to five-year (i.e., multitemporal NLCD) land cover maps generated on an approximate five-year interval? We used the Soil and Water Assessment Tool SWAT [15,16] to model flows using both annual (land cover maps that include 22 years of disturbance from the National Forest System (NFS)), and five-year (land cover maps based on multitemporal NLCD products) across watersheds of the Intermountain West. We used SWAT due to its versatility with respect to geospatial inputs, ability to generate multiple scenarios under different land cover and climate regimes, and adaptability to work in agricultural [17], tropical [18], and snowmelt-driven [19] watersheds. Here, we combine highly-detailed disturbance datasets with a physically-based hydrologic model to predict water dynamics in 10 forested watersheds. Each of the watersheds is within the US NFS, which manages 75 million hectares of public land. A detailed map of forest disturbances was created across the NFS using Landsat imagery and conventional change detection and image interpretation methods. Our objective—to assess any benefits in predicting water yield by using annual land cover data—is a starting point for leveraging recent advances in satellite time series analysis [20] in the monitoring and prediction of important elements of the forest hydrologic cycle. 2. Materials and Methods 2.1. Study Area We selected 10 watersheds of different sizes and ecological characteristics across the Northern US Rocky Mountains (Figure 1). The criteria for selecting watersheds to be used as pilots were as follows: (a) the watershed’s spatial extent should be completely contained within the spatial domain of disturbance datasets developed across the NFS—this would guarantee a comprehensive understanding of disturbance in these watersheds; and (b) within the watershed, there should be at least one hydrometric station with sufficient (at minimum, five years) flow records that could be used to assess the goodness of fit of SWAT’s monthly flow outputs. A selected number of landscape and climatic characteristic for the watersheds in the pilot study are summarized in Table 1. The attributes of the USGS hydrometric stations that were used for calibration and validation of the SWAT model in each watershed area were also provided. Notice that six stations had complete time series for the discharge data, and four stations had partial coverage for the 1990–2011 time series. Therefore, model performance for units 07, 10, 13, and 14 when reported only covers those years that overlap this study’s time series. Links to all stations used in this study are provided in Table S1 in the Supplementary material. The daily stream gauge discharge data are reported in cubic feet per second, so in order to use it with SWAT’s outputs, the data had to be converted to cubic meters per second. Table 1. Pilot watersheds characteristics and hydrometric stations utilized for model performance Unit
Area (km2 )
EL (m) Range
FTG > 25% †
SHG ‡
Pcp (mm) ⊗
T (◦ C) ⊗
USGS Code
Long.
Lat.
Years of Record
Wat01 Wat02 Wat03 Wat04 Wat05 Wat06 Wat07 Wat10 Wat13 Wat14
198.3 50.1 3026.0 855.0 1717.2 3360.9 177.2 62.7 2552.7 941.3
2496–4029 1912–3591 1096–2838 1141–2787 442–2445 509–2407 2435–3924 2284–3484 671–2850 660–2598
FSMH–LP PJ–AB FSMH–LP DF–FSMH DF–FSMH FSMH–DF FSMH–LP FSMH–LP DF–FSMH FSMH
B–C B B–D B B–C B–C B B B–C–D B–C–D
899.4 536.8 1145.2 990.4 1019.5 1401.8 984.3 1039.9 1061.6 1248.3
0.1 4.2 2.5 3.2 5.2 4.9 0.5 1.9 3.6 4.1
09289500 10249300 12359800 13310700 13336500 13340600 9277800 10153800 13335690 13335700
40◦ 360 24” 38◦ 530 15” 47◦ 580 44” 44◦ 590 13” 46◦ 050 12” 46◦ 500 26” 40◦ 330 27” 40◦ 350 48” 46◦ 070 18” 46◦ 070 28”
110◦ 310 35” 117◦ 140 40” 113◦ 330 38” 115◦ 430 30” 115◦ 300 50” 115◦ 370 16” 110◦ 410 50” 111◦ 050 48” 114◦ 550 50” 114◦ 550 58”
1933–2017 1965–2017 1965–2017 1966–2017 1911–2017 1967–2017 1965–1994 1963–1996 1995–2006 1995–2005
† Predominant Forest Type Group FTG [21]: FSMH: Fir/Spruce/Mountain Hemlock Group; LP: Lodgepole Pine Group; PJ: Pinyon/Juniper; AB: Aspen/Birch Group; DF: Douglas-fir Group. ‡ Soil Hydrologic Group SHG. ⊗ Pcp (mean annual precipitation), /T (mean annual temperature)—obtained from zonal statistics analyses using PRISM [22].
Forests 2018, 9, 317 Forests 2018, 9, x FOR PEER REVIEW
3 of 19 3 of 19
In the table above, the areas of the the different different forest forest type type groups groups (FTG) (FTG) were were calculated calculated for each watershed by intersecting the FTG geospatial layer with each watershed’s boundaries. watershed by intersecting the FTG geospatial layer with each watershed’s boundaries. The first and second largest FTG are listed listed for for each each watershed. watershed. The FTG dataset and corresponding metadata for the contiguous contiguous US and Alaska can be accessed accessed from the the US US Forest Forest Service Service Geodata Geodata Clearinghouse Clearinghouse (https://data.fs.usda.gov/geodata/rastergateway/forest_type/). One of the key parameters that hasa (https://data.fs.usda.gov/geodata/rastergateway/forest_type/). One of the key parameters that has abig biginfluence influenceon onSWAT’s SWAT’sstreamflow streamflowoutputs outputsisisthat thatof of the the soil soil hydrologic group (SHG). These are classified into four categories in a spectrum—A, B, C, and D—based on infiltration characteristics classified into four categories in spectrum—A, characteristics and surface runoff conditions that are required to simulate the movement of water in the soil, and the corresponding corresponding conveyance conveyance to to the the reach. reach. Soils type A have the best characteristics for high infiltration and low runoff whereas soils type areare thethe opposite, andand B and C are intermediate classifications [23]. low runoff whereas soils typeD D opposite, B and C are intermediate classifications The of theofsoil groups in each was obtained following the same [23].predominance The predominance thehydrologic soil hydrologic groups inwatershed each watershed was obtained following the procedure as the as FTG. same procedure the FTG.
Figure 1. 1. Location Location of of the the10 10pilot pilotwatersheds watershedsininthe thecontext contextofofthe the conterminous US. The inset shows Figure conterminous US. The inset shows in in detail distribution in the Intermountain West. detail thethe distribution in the Intermountain West.
2.2. The The SWAT SWAT Model Model 2.2. SWAT(the (theSoil Soil and Water Assessment haswidely been widely for theofanalysis of the SWAT and Water Assessment Tool)Tool) has been used forused the analysis the hydrologic hydrologic impact of land cover and management in tropical and temperate regions. The following impact of land cover and management in tropical and temperate regions. The following basic basic description the model obtained from [24–26]. For modeling,SWAT SWATdivides dividesthe thebasin basin into into description of theofmodel was was obtained from [24–26]. For modeling, sub-basins, and these into hydrologic response units (HRU), consisting of homogenous areas of soils, sub-basins, and these into hydrologic response units (HRU), consisting of homogenous areas of soils, land cover/use, cover/use, and land and topography topographythat thatcan canimpact impactthe thehydrology hydrologyof ofthe the study studyarea area in in aa similar similar way. way. The SWAT SWATmodel model simulates what happens the by basin by predicting each component of the The simulates what happens in thein basin predicting each component of the hydrologic hydrologic cycle. The main are two divided intoterrestrial two phases: andAhydrological. A cycle. The main processes areprocesses divided into phases: and terrestrial hydrological. brief summary brief summary of both phases is provided as supplement material. of both phases is provided as supplement material.
2.3. SWAT’s Inputs and Principal Processes In this study, we used ArcSWAT version 2012.10_2.18 [27] within the ArcMap 10.2.2 environment. ArcSWAT is freely available, and ArcGIS version-specific ArcSWAT extensions can be obtained from
Forests 2018, 9, 317
4 of 19
the US Department of Agriculture–Agricultural Research Service website (https://swat.tamu.edu/ software/arcswat/). 2.3.1. Watershed Delineation Delineation of the pilot watersheds was based on a 30 m grid cell size digital elevation model (DEM) obtained from the Geospatial Data Gateway [28]. A source area delineation threshold [29] was adopted for each pilot watershed. Following this criterion guaranteed that the average size of delineated sub-watersheds within each watershed would be an average of 4.5% of the delineated basin, which has been suggested [29,30] as the minimum threshold to detect the impact of management practices. In this context, each pilot watershed had an average of 22 sub-watersheds. The outlet location for each pilot watershed was determined by the coordinates of the hydrometric stations (Table 1) used to assess the model’s goodness of fit. 2.3.2. Climate Data Daily summaries of total precipitation and air temperature (maximum, minimum, and mean) for all the pilot watersheds were obtained from the National Centers for Environmental Information [31]. A period of record dating back to the five years before the beginning of our time series (1985), and ending with the most recent records (2016) was defined, hence completely including the time period of our disturbance datasets (1990–2011). Hydrologic cataloging units (HUC levels 08 and 12) were used as spatial domains to select all available climatic data. All stations available at the HUC level 08 (even if they were located outside the actual boundaries of the watersheds) were initially selected. This provided the best observed climatic data within each watershed or in closest proximity to it. However, only those climate stations with a temporal coverage of at least 90% of our time series were ultimately chosen. An advantage to using this climate source is that data are already provided in the proper units (metric), and in the format required for SWAT analyses. Table S2 is available in the Supplementary Material. This table contains the name of the HUC level 08 that each watershed belongs to, as well as weblinks where the daily climatic records can be accessed. We defined elevation bands for each pilot watershed to account for elevation gradients and their effects on snow accumulation and melt processes [30] in mountainous watersheds. Three elevation bands were used in each pilot area. 2.3.3. Land Cover Characteristics The focal point of this paper is to determine the effects, if any, of using alternative land cover maps with SWAT. The two alternatives were: five-year (multitemporal NLCD), and annual (integration of disturbance products) maps. SWAT models were fit to each watershed keeping all other inputs (DEM, climate, and soils—described below) and SWAT default parameters constant. A description of each land cover set follows. Whenever five-year maps are mentioned throughout the paper, it means that these land cover maps were refreshed or updated with a five-year delay instead of annually, which is the case for the annual maps.
•
Five-year land cover maps
All available (1992, 2001, 2006, and 2011) multitemporal NLCD products were used in this step. Our assumption here is that the 1992 product resembles the land cover conditions at the time when our disturbance datasets start (i.e., 1990), and then the subsequent products (2001, 2006, and 2011) can be used to describe changes in the landscape that have occurred for the complete duration of our time series. No changes were made to these maps in this stage. Nevertheless, SWAT only ingests a single description of the land cover conditions for the entire duration of a regular simulation. In other words, it assumes by default that land cover does not change for whatever number of years are included in the simulation. Even though SWAT users are able to utilize NLCD products during simulations without adjustment, we needed to establish a process to make both datasets (annual and five-year maps) directly comparable in order to guarantee that no advantages were given to one dataset over the
Forests 2018, 9, 317
5 of 19
other. The process to integrate the four available temporal NLCD scenarios in one SWAT simulation is described directly after the description of the annual land cover.
•
Annual land cover maps
A long-term annual (30 m spatial resolution) time series was created that showed the timing, extent, and type of disturbance across the NFS from 1990–2011. As a first step in the production of these maps, an automated forest change detection algorithm [32] was run on an annual, cloud-free series of Landsat TM and ETM+ imagery. Technicians manually edited this map by adding, subtracting and changing the shape of disturbance patches using several sources of reference data, including: multi-temporal composites of Landsat data transformed to one-band per year with the Disturbance Index [33]; high-resolution time series of aerial imagery served through Google Earth; a spatial database of historical harvest activities (the National Forest System’s Forest service Activity Tracking System or, FACTS); a national interagency database of fires over 405 hectares in size [34]; and insect activity records obtained from aerial detection surveys [35]. This process also allowed the technicians to label the type of disturbance (harvest, fire, bark beetle) for each affected patch.
•
Fusion of a forested land cover baseline with disturbance maps
For SWAT models using both five-year and annual land cover maps, we utilized a product that describes the spatial distributions of forest types groups (FTG) as described in [21]. It has been demonstrated [30] that explicitly integrating the configuration of the forested landscape (i.e., which pixels are Ponderosa Pine forests, which pixels are Douglas-Fir forests, etc.) in SWAT provides much better results. The FTG geospatial dataset was used as baseline for the 1990 year for both datasets. The five-year maps were spatially-intersected with the FTG dataset, and the forest spatial domain from the FTG was maintained. Pixels from the five-year maps that did not exactly match the FTG spatial domain were assigned to a FTG class based on proximity. This process was repeated for all the five-year maps. In this context, the parameters for the NLCD forest classes were not used in SWAT. Rather, the specific parameters for FTG classes were used. This cross-walking (NLCD to FTG) activity guaranteed that the five-year maps were using the same SWAT parameters that the annual maps used. In the case of the annual maps, and starting with the 1990 FTG map, each subsequent year was then updated to reflect the changes that occurred on the landscape due to the mapped disturbances. For instance, the land cover conditions for the very next year (1991) were determined to be the same FTG from 1990 with the exception of the disturbed pixels, which were recorded as a new land cover class that we named ‘disturbed’. After another year, the classification was allowed to go back to ‘forest’ and the original FTG class. This is, in essence, the difference between the annual and five-year maps: the periodic use of the disturbed class. We had considered keeping disturbed pixels as disturbed in subsequent years but did not have the data to determine what the land cover condition was post disturbance. This is a limitation in our analysis. Parameters for these forest classes, and the disturbed class with regards to interception capacity, canopy density, and surface roughness were obtained from the application of SWAT in snowmelt-driven catchments reported by [30]. 2.3.4. Soil Characteristics We used the State Soil Geographic database STATSGO [23] to characterize the effects of different types of soils in the hydrologic regime of each watershed. These datasets describe general soil properties and contain the required attributes for SWAT to simulate the movement of water in the soil, and the corresponding conveyance to the reach. While the Soil Survey Geographic (SSURGO) [36] has a more detailed spatial resolution than that of STATSGO we did not find complete coverage for all of our pilot watersheds. According to [37] SWAT models fitted using STATSGO produced better accuracies for low streamflows and better performance during validation. The predominant soil hydrologic groups (SHGs) (i.e., A, B, C or D) [38] for each watershed are listed in Table 1.
Forests 2018, 9, 317
6 of 19
2.3.5. Hydrologic Response Units Hydrologic response units (HRUs) are intrinsically the main modeling unit in SWAT. Each watershed is divided into unique combinations of land cover/FTG, soils, and management practices prior to conducting a simulation. Due to their spatial homogeneity, uniform hydrologic responses are expected from each HRU. There are several options in SWAT which control the spatial variability and number of HRUs, which can influence the outputs [39] obtained during simulations (i.e., a greater number of HRUs capture more variability). In this study, we defined a threshold of 0% minimum of watershed area for land cover and soils units, thereby keeping all possible combinations of land cover and soils. This guaranteed a high level of resolution for modeling purposes. More details are provided below. A single slope class was used during HRUs definition, but as mentioned above the effects of snow accumulation and melt processes were included during the modeling by defining three elevation bands, which has been found [40] to improve SWAT’s performance in mountainous terrain with snowmelt-driven hydrology. 2.3.6. Integration of Land Cover Change in SWAT With the implementation of SWAT version 2009, a new module was added to account for land use and land cover changes that take place over the time series or the modeling period. Before this module was added, an assumption that the land cover in a given watershed did not change during the modeling period was made. This assumption is usually not true, especially during modeling periods that span several decades, such as the ones conducted in this study. A cloud-based software tool (SWAT LUU) can ingest multiple land cover geospatial datasets and produce the required inputs for the SWAT land cover change module. The multiple five-year (1992, 2001, 2006, and 2011) temporal datasets (modified as explained above through fusion with a forested land cover baseline), and the annual land cover maps (1990–2011) were uploaded to SWAT LUU [41] in conjunction with their corresponding SWAT modeling projects to obtain the required inputs for simulation of land cover change. This process was conducted for each one of the pilot watersheds (i.e., one project using the five-year maps, and another project using the annual maps). With each new land cover map that is available, SWAT LUU recalculates the proportions that each HRU (i.e., Ponderosa pine with an ‘x’ soil series, disturbed forest with a ‘y’ soil series, etc.) occupies in the watershed under study. With an updated description of the landscape, and with the changing climatic conditions, SWAT is then able to model the effects of land cover change. Because SWAT cannot use land cover classes not present in the first year of the simulation, a negligible amount of ‘disturbed’ land (five 30-m pixels per soil group) were added to the FTG map in 1990. This process was required for both annual and five-year maps. 2.3.7. SWAT Simulations For each pilot watershed, two simulations were set up in SWAT: one using the five-year maps datasets and the other using the annual land cover maps which include the disturbance time series. All inputs (DEM, climate, soils), and parameters influencing those inputs were kept the same for the two simulations with the exception of land cover. Default parameters that relate to canopy structure, interception capacity, and surface roughness for the simulations using the annual maps were modified following [30] the forest classes (Lodgepole, Spruce-fir, and disturbed forest), and for the rangeland and barren classes. The Ponderosa and Douglas-fir forests classes were adjusted using a combination of the characteristics detailed by [30], and an assessment of leaf area index time series [42–44]. Tree growth default parameters for both simulations (five-year and annual land cover) were also modified. Preliminary runs of these models indicated that transpiration was much lower than expected. As pointed out by [45], this was due to a default operation practice in SWAT that kills and harvests the crops or forests every year. Such a forestry operation was not expected in our watersheds, and thus was removed, and the initial settings for tree growth for both types of simulations (five-year and annual) were modified accordingly.
Forests 2018, 9, 317
7 of 19
2.3.8. Model Calibration Once the initial models were fitted as described above, we conducted a calibration process so that the SWAT parameters used during model validation would be based on the local specific conditions of the 10 watersheds, and not on SWAT’s default parameter values. This is particularly important given that all of our watersheds’ hydrologic regimes are snowmelt dependent, and SWAT’s default parameters are specific for agricultural watersheds. SWAT’s parameters are process-based and must be used within a sensible and realistic range that can change from watershed to watershed. The main goal of the calibration process is to determine which parameters have a predominant influence on the hydrology of the watershed, and to estimate a range of values for those parameters that better reflect the physical processes while at the same time reduces prediction uncertainty. An initial set of 26 parameters that have been previously documented [30] as being significant in affecting the hydrology of this type of watersheds was used. The list of parameters, and the initial range of values are provided as supplement material. Model calibration was performed using the Sequential Uncertainty Fitting (SUFI2) semi-automated algorithm that is available within the SWAT Calibration and Uncertainty Programs SWAT-CUP software [46]. In SUFI2 many iterations are performed over the selected subset of parameters, and the initial range of values. As more iterations are applied, the “parameter ranges get smaller zooming in a region of the parameter space, which produces better results than the previous iteration” [47]. The user can tune the uncalibrated model by minimizing the differences between observed and simulated flows. This process can be done by defining an objective function to optimize, which in this study it was the Nash–Sutcliffe efficiency (NSE) technique. The NSE is “a normalized statistic that determines the relative magnitude of the residual variance” [48], and indicates how well the plot of observed versus simulated data fits the 1:1 line [49] A thorough description of calibration and validation concepts, and applications can be found in [47]. The calibration was conducted independently for each of the pilot watersheds, however the initial set of parameters to be calibrated was the same for all watersheds. The calibration for each watershed was conducted as follows: (a) with the initial set of parameters and their initial values Table S3, 750 simulations were run to obtain the results for the first iteration; (b) the calibration outputs—i.e., the global sensitivity analysis, of the iteration were reviewed to determine which parameters should not be included in the next iteration. We chose to leave out those parameters (usually one or two) with the highest p-values, which may be interpreted as not having a significant impact in changing the predicted flow values; (c) we then reviewed the new parameter range values after the iteration, and copy those values for the parameters that we decided to include in the next iteration; (d) a new iteration was run with the chosen subset of parameters and their new values using 750 simulations; (e) the process to select the new subset of parameters to be used in the next iteration was repeated, and the change in NSE was noted; (f) this process was repeated until there was not a visible and noteworthy change in NSE improvement from iteration to iteration; (g) the SWAT original model was modified accordingly using the final subset of parameters, and their calibrated values. This is a computationally demanding process, and SWAT-CUP had to be used with an extension that allows parallel processing, thereby reducing the required time to run hundreds of simulations. 2.3.9. Model Validation The accuracy of the SWAT calibrated models using the annual and five-year land cover was assessed independently. The NSE technique [48], the percent bias (PBIAS) measure [50], and the ratio of the root mean square error to the standard deviation (RSR) were calculated by comparing the simulated SWAT flows with the observed monthly statistics (monthly average of flow discharge) at each one of the watersheds outlets. These averages were obtained directly from the time series: monthly statistics that have been calculated for each hydrometric station and that are available from the USGS website (https://waterdata.usgs.gov), Table 1 and Table S1. During SWAT’s model configuration, the spatial location of these outlets was defined by using the same coordinates of USGS’s surface–water hydrometric stations. The performance of SWAT was estimated for either the entire disturbance
Forests 2018, 9, 317
8 of 19
time series (1991–2011) or just for the years that data were available at each hydrometric station and overlapped our period of analysis. All stations had complete monthly availability, except Wat03 for which no records were measured for the months of December through March, thus the performance that is reported for this watershed is based only on eight months of data per year. The characteristics of these stations are provided (Table S1) as supplement material. NSE has been recommended [49,51] in the literature as the standard to compute hydrologic model performance. NSE is an excellent indicator of how well the simulated data mimics the observed data and ranges from −∝ to 1.0. Values < 0 are thought to indicate an unacceptable performance, and values > 0 and
